Dynamic scene modelling and anomaly detection based on trajectory analysis

Abstract

A real-time scene modelling approach is presented that recognises temporary and permanent road structure change resulting from construction, accident or lane expansion and other obstructions. The system defined utilises a two-phase approach to modelling the scene. In the transitional phase, a dominant set-based graphical clustering approach is applied to understand the current scene structure from trajectory groupings, whereas the operational phase analyses the trajectories in real-time to detect anomalies such as u-turns, wrong-way or erratic drivers based on the acquired model of the scene structure and normal traffic patterns. In addition, the concept of dynamic traffic flow analysis is utilised to identify and remember temporary additions and removals of paths due to construction and accidents, as well as permanent road structure changes. An intuitive equal-arc-length sampling is applied to extract only the spatial information from the trajectory comparisons, since the spatial characteristics alone are sufficient for road structure understanding. A distance metric is developed to measure spatial difference and directional change of the path with entrance and exit awareness. Results for a publicly available dataset are provided, demonstrating that the method can efficiently model the scene, detect anomalies and capture both temporary and permanent scene reconstructions.